bergen, norwayvol. i · no. 20 · July 17, 2026rss feed

Hasan Arief

A lab notebook on agentic coding, open-weight models, and what they cost to run

Uses

The working stack behind these notes, kept current as it changes. Nothing here is sponsored; most of it is interchangeable.

Coding agents

Claude Code as the primary interactive driver, with hooks, skills, and subagents doing the safety and review work. Codex CLI for unattended mechanical batches and as a second opinion on risky changes. Both read one instruction file; the reasoning behind the split is inthe comparison guide.

MCP servers

Two earn daily use: Playwright MCP for browser verification and mobile-mcp for real-device app testing. Thefull note covers why the rest lost to plain shell tools.

Research stack

PyTorch and scikit-learn for modelling; ObsPy and SciPy for seismic and acoustic signal processing; HPC clusters for training and ensemble runs. For LLM tooling work: locally hosted Mistral for text, Qwen Vision for multimodal input, wired through MCP tool catalogs with typed outputs.

This site

Astro with MDX content collections, zod-validated frontmatter, static output served by nginx behind a CDN. Every entry ships as a pull request through CI gates (build, links, JSON-LD, Lighthouse, and a style check) and a human review. The setup is documented inthe setup guide.

Updated in place; last revision 2026-07-04.